prosody-syntax integration in a second language

Second Language Research 1 –28

© The Author(s) 2016Reprints and permissions:

sagepub.co.uk/journalsPermissions.navDOI: 10.1177/0267658316649998

slr.sagepub.com

secondlanguageresearch

Prosody-syntax integration in a second language: Contrasting event-related potentials from German and Chinese learners of English using linear mixed effect models

Stefanie Nickels and Karsten SteinhauerMcGill University, Canada; Centre for Research on Brain, Language and Music (CRBLM), Canada

AbstractThe role of prosodic information in sentence processing is not usually addressed in second language (L2) instruction, and neurocognitive studies on prosody–syntax interactions are rare. Here we compare event-related potential (ERP) data of Chinese and German learners of English L2 to those of native English speakers and show how first language (L1) background and L2 proficiency influence the online processing of prosody-induced garden-path effects. Unlike most previous ERP studies, we use linear mixed effect models to analyze L2 proficiency as a continuous (rather than categorical) variable. Our results extend previous findings and provide unique data on complex interactions between L1 background, prosodic structure, and morphosyntactic processes in real time. The reader is introduced to this new approach and also learns why coverage of prosody in L2 instruction may be beneficial.

KeywordsClosure Positive Shift (CPS), event-related potentials (ERPs), language proficiency, language teaching, language transfer, linear mixed effect models, prosodic boundary, syntax

I Introduction

Prosody, i.e. the rhythm and intonation of speech, is one of the fundamental aspects of language. While it is consistently present in speech and affects both production and com-prehension, it is rarely, if ever, part of a second language acquisition syllabus. In addition to conveying information about the speakers’ affective status (e.g. joy, anger), prosody is

Corresponding author:Stefanie Nickels, School of Communication Sciences and Disorders, McGill University, 8th Floor, 2001 McGill College, Montreal, Quebec H3A 1G1, Canada. Email: [email protected]

649998 SLR0010.1177/0267658316649998Second Language ResearchNickels and Steinhauerresearch-article2016

Special Issue on Neurolinguistics

at MCGILL UNIVERSITY LIBRARY on June 30, 2016slr.sagepub.comDownloaded from

mailto:[email protected]

http://slr.sagepub.com/

2 Second Language Research

often essential in conveying the correct interpretation of an utterance. This is the case for sarcasm and irony, for distinguishing statements (this is edible!) from questions (this is edible?), and also for many syntactic ambiguities that can be effectively resolved by prosodic phrasing (Kjelgaard and Speer, 1999). The boundaries of prosodic phrases are acoustically marked by a number of cues, including the lengthening of the clause-final word, characteristic pitch changes (especially boundary tones) and the insertion of a pause (Steinhauer et al., 1999).

Large prosodic boundaries often coincide with major syntactic boundaries. This aspect is what makes them so valuable for the listener, because they serve as a salient guide to the underlying syntactic structure of a sentence. Take for example the sentence When a bear approaches the people come running. As predicted by the Late Closure (LC) princi-ple (Frazier, 1987), initially the subject noun phrase the people tends to be (mis-)inter-preted as the object of the preceding verb approaches, resulting in longer processing times (‘garden-path’ effect). A prosodic break (or a comma) before the people prevents this initial misunderstanding by closing the verb phrase for additional words (Early Closure or EC). Since prosodic boundaries can determine the initial syntactic analysis and are difficult to ignore (or to ‘mentally delete’), an overt boundary in the wrong position has been found to cause even more severe processing problems than a missing boundary (Bögels et al., 2013; Pauker et al., 2011). Compared to prosodic boundaries, other types of lexical, syntactic, or contextual information may also modulate EC/LC parsing decisions; however, their impact is usually weaker. For example, if verbs are preferably used in transitive structures, the LC bias is stronger than for intransitively biased verbs (Itzhak et al., 2010). Moreover, while the LC preference tends to be strong for verb phrases in past tense, progressive aspect seems to reduce the need for a direct object (compare I was eating and *I ate), thus substantially reducing the LC bias (Frazier et al., 2006).

Prosodic support of syntactic structures has been shown to be implemented naturally and frequently in spoken language (Schafer et al., 2000), yet it is not covered in the tra-ditional classroom instruction. In fact, if prosody is taught in the classroom at all, it seems to be mostly restricted to pronunciation, especially word stress (Mennen, 2007; Saito, 2012; Trouvain and Gut, 2007). In a survey of prosody-related articles published in second language (L2) research journals for the past 25 years, Gut found only four articles on the perception of intonation patterns (Trouvain and Gut, 2007), as well as a few conference proceedings; e.g. Hwang and Schafer (2006) on L2 boundary processing by Korean learners of English. The present investigation addresses precisely this under-researched area in L2 learning. Specifically, we focus on how prosodic boundaries influence syntactic parsing, and how well second language learners from different L2 backgrounds can use these cues to aid their understanding of syntactically ambiguous sentences. Following Pauker et al. (2011), our experimental EC and LC sentences were constructed such that prosodic boundaries either cooperated or conflicted with the under-lying syntactic structure of the sentence (see Table 1).

II Research questions

Using these English auditory stimuli, we compared German and Chinese native speak-ers to English native speakers. The two groups of second language learners differ considerably in their first language’s (L1) similarity to English. German and English are



Nickels and Steinhauer 3

both Germanic intonational languages, and they share many cognates and the same alphabet. Chinese, on the other hand, is part of the Sinitic language family, has a logo-graphic writing system, and uses various lexical tones to distinguish between different lemmas. As both German and Chinese speakers use intonational phrasing in their mother tongues, one might predict only marginal group differences for their prosodic processing in English; however, we are not aware of any previous study comparing the two groups. Moreover, there are a number of specific research questions. First, both German and Chinese learners of English tend to have problems with tense, and even more so with aspect in English (e.g. progressive aspect expressing an ongoing/repeated action as in she was sleeping/jumping vs. she slept/jumped), because both languages differ consider-ably from English in terms of their aspect levels as well as the way aspect is expressed1 (e.g. Dürich, 2005, for German; Shi, 2011, for Chinese). As our English sentence mate-rial uses either present or past progressive for the first verb to reduce the strong LC bias seen in simple present or past tense (compare Frazier et al., 2006) in native speakers (e.g. When a bear is approaching the people …), one might predict stronger LC biases in L2 learners who are less used to progressive aspect. If so, L2 groups should show a stronger impact of the first prosodic boundary disambiguating towards EC.

Second, whereas both German and English listeners have been found to show stronger garden-path effects with superfluous boundaries (in condition D) than with missing boundaries (condition C) (Nickels et al., 2013; Pauker et al., 20112), it is unclear whether this pattern predicted by the Boundary Deletion Hypothesis also generalizes to Chinese listeners.

Third, there is an ongoing debate about whether late L2 learners have to rely on pro-cessing mechanisms distinct from native speakers (e.g. Bley-Vroman, 1989; Lenneberg, 1967; Weber-Fox and Neville, 1996) or can converge on the real-time processing seen in native speakers (e.g. Steinhauer et al., 2009). We will address these issues by including language proficiency as a factor. Although it is quite obvious that language proficiency must play a role in second language processing, quantifying second (and first) language proficiency can be difficult. While there are many proficiency measures available, they will most likely fail to capture the exact skills needed to process the specific linguistic structures used in a given study. For this reason we adopted a strategy suggested by Steinhauer et al. (2009), namely to use the structure-specific proficiency as revealed in

Table 1. Example sentences.

Condition Part 1 Splicing point Part 2

A (LC) When a bear is approaching the people # the dogs come running

B (EC) When a bear is approaching # the people come running

C (A1–B2) When a bear is approaching the people come running

D (B1–A2) When a bear is approaching # the people # the dogs come running

Notes. Conditions C (without a boundary) and D (with two boundaries) are hybrid versions of conditions A and B, derived by cross-splicing. LC = Late Closure; EC = Early Closure, # = boundary.




the by-participant accuracy at the end-of-sentence judgments (‘Does it sound natural or not?’) as the operationalization of proficiency.

In order to track the online processing patterns in our three groups, we use event-related brain potentials (ERPs). Successful processing of prosodic boundaries elicits the Closure Positive Shift (CPS), a positive-going waveform near the midline electrodes that has been replicated in many languages (for a review, see Bögels et al., 2011). Difficulties resulting from prosody-syntax mismatches in garden-path sentences are reflected by N400 and P600 components at disambiguating elements later in the sentence (Pauker et al., 2011). The N400 is a negative-going brain wave most strongly associated with lexical-semantic processing difficulties, including those involving theta role assign-ments, whereas the P600 is at least partly linked to processing costs during structural integration (e.g. Steinhauer and Connolly, 2008). In condition C, a relatively small P600 reflecting the structural (and prosodic) reanalysis of a rather weak garden path was previ-ously found on the disambiguating second verb phrase come running (Pauker et al., 2011). In condition D, the second boundary and the onset of the third noun phrase (the dogs) elicited first an N400 around 400 ms (reflecting problems during lexical integra-tion and theta role assignment of the prosodically detached NP2 the people) and then a large P600 (indicating the processing demands of a severe garden path effect). In L2 research on grammar processing, these ERP components tend to be more robust, earlier and larger with increasing L2 proficiency, and often converge on L1 patterns once native-like proficiency has been attained (Steinhauer, 2014).

In sum, this ERP study investigates the influence of L1 background and proficiency with respect to prosody-syntax integration. Our research question is two-fold with a sup-plementary question:

1. How do L1 background and proficiency influence the processing of a prosodic boundary (i.e. using correct, non-anomalous sentences)?

2. How do L1 background and proficiency affect the integration of prosodic and syntactic cues in sentences where those two sources conflict with each other?

3. Lastly, which recommendations can we infer from those results for language teaching?

III Methods

1 Materials

The four experimental conditions are displayed in Table 1. The first part of each sentence contained an optionally transitive verb that was either followed by an object as in A (Late Closure), or remained intransitive as in B (Early Closure). Only sentences A and B had been recorded by a male Canadian English native speaker, while sentences C and D were created by carefully cross-splicing the first and second half of sentences A and B at the fricative ‘th’ (θ) (for a detailed description of the material creation, see Pauker et al., 2011). This procedure is preferable to recording the violations directly because it allows the materials to be acoustically identical, while at the same time avoiding a potentially unnatural pronunciation due to the violation.




Note how the recombining of the sentence halves results in violation condition C, which is missing the disambiguating boundary after approaching. This missing boundary facilitates an incorrect interpretation of the people as the object of the first verb phrase, when it is in fact the subject of the next clause. This garden-path phenomenon becomes apparent upon hearing come running. Condition D, on the other hand, carries a superfluous boundary, which incorrectly signals an early closure of the first phrase after approaching, when in fact the people is the object of that sentence, and should thus syn-tactically be grouped with the first phrase. This conflict becomes apparent when the second boundary is encountered. A total of 160 sentences (40 sentences per condition) were presented, pseudo-randomly intermixed with 160 phrase-structure violations serv-ing as filler sentences (e.g. The man hoped to *meal the *enjoy). It is possible that those fillers were more salient to the participants than the prosody-induced violations. However, this potential effect is consistent for all groups and conditions in this study, and is thus unlikely to have contributed to differences between groups or conditions.

2 Participants

All participants were dominantly right handed (Oldfield, 1971), had normal or corrected-to-normal vision and reported to have had no previous neurological disorders. Out of the 105 tested participants, 89 were included into the analysis after artifact rejection (50 female, 39 male). The included participants were made up of 20 English, 39 German, and 30 Chinese speakers (27 Mandarin, 3 Cantonese). The study received Institutional Review Board (IRB) approval from all universities involved, and participants gave writ-ten informed consent. All German participants were tested at Saarland University, Germany. English native speakers and the majority of Chinese speakers were tested at McGill University, Canada, while eight of the Chinese speakers were tested at Bangor University, UK.

3 Procedure

Participants were seated in a comfortable chair in front of a computer monitor inside a shielded, sound-attenuating booth (IAC Inc.). Their task was to judge the acceptability of each sentence using the computer mouse (‘Natural or not?’), while their electroencepha-logram (EEG) was recorded. The session began with a short practice block containing 10 sentences.

4 Data recording and preprocessing

EEG data were recorded from 19 Ag-AgCl electrodes (10-20 system) at a 500 Hz sam-pling rate, in Montreal and Bangor using a Neuroscan Synamps2 amplifier, in Germany using a BrainProducts BrainAmp amplifier. Data were preprocessed using EEGLAB (Delorme and Makeig, 2004) and ERPLAB (Lopez-Calderon and Luck, 2014). After import, the continuous EEG was re-referenced to the right mastoid, and filtered using ERPLAB’s IIR Butterworth filter. Band-pass cut-off frequencies were .1 and 30 Hz (−6 dB). The filter was specified as a second order filter resulting in a gradual roll-off




(12 dB/octave). After epoching the data to −500 to 1500 ms around the onset of the boundary (CPS) or the splicing point (garden path effects), data were rejected if (1) the eye channels exceeded a peak-to-peak threshold of 75 µV within moving 100 ms time windows (100 ms steps), or if (2) any scalp channel exceeded an absolute threshold of ±75 µV during the entirety of the epoch. This procedure led to the exclusion of 16 partici-pants for whom more than 50% of the trials were rejected. The remaining 89 participants had a mean of 18% rejected trials (range: 0%-46%, median: 13%).

5 Statistical analyses

This study aimed at simultaneously evaluating the effects of L1 background and structure-specific L2 proficiency (see Steinhauer et al., 2009). A good operationalization of the latter would be a d prime (d′) measure contrasting hits and false alarms across all correct and violation conditions. However, EC condition C (missing boundary, but progressive tense) was found to be acceptable for native speakers in about 50% in a previous study (see Pauker et al., 2011), and therefore does not qualify as a clear violation condition. As a small d′ value in the B vs. C contrast may not reflect low language proficiency, we decided to calculate d′ only for the contrast between condition D (clear violation due to a superfluous boundary) and its lexically matched control condition A. Note that the d′ measures discriminability between acceptable and unacceptable structures while penal-izing response biases (such as to predominately reject a sentence when in doubt). It also correlated positively with two external measures of proficiency. First, a subset of 46 of the L2 speakers were administered part of the Cambridge Certificate of Proficiency, which was positively associated with d′ (r(44) = 0.46, p < .01). The participants’ self-assessed L2 proficiency also correlated positively with d′ (r(68) = 0.38, p < .01). These positive cor-relations suggest that the structure-specific d′ is a good operationalization of language proficiency for the current study. These d′ scores by group can be seen in Figure 1.

Examining the different groups on their d′ score reveals a challenge: the Chinese group has a lower score, while Germans and English speakers do not differ from one another. This means that group and proficiency are confounded as the Chinese group always shows the lowest range of proficiency. This confound is a common problem in second language research. Some researchers have opted to either use a hierarchical regression, or a residualized predictor to address this issue, although they result in what is likely an undesired effect: the effect of the residualized variable stays the same, while the result for the other variable changes (for details see Wurm and Fisicaro, 2014). For this study, we thus chose to enter the two collinear variables as they are, but to interpret the results specifically for the Chinese group with caution.

Further, the inclusion of the native speaker group in this analysis allows us to assess the influence of proficiency across all groups. This is warranted since the native speakers also show a considerable amount of variation in their answering patterns (see Figure 2), and previous behavioral and ERP studies also showed that native speakers differ in their L1 proficiency (Alderson, 1980; Pakulak and Neville, 2010).

It is important to note that age of acquisition (AoA) is also often discussed as a deci-sive factor in second language processing. For the current study, AoA is also particularly confounded with group, as all English native speakers have an AoA of zero, and the




Figure 1. d′ for the discrimination ability between A and D by group.Note. A d′ of zero indicates no discrimination ability.

Figure 2. Percentage of accepted trials by condition and group.Note. The horizontal line within the box is the median of the distribution; the number is the mean.




German native speakers show an AoA between 10 and 14 (mean: 11.8), the typical onset of English instruction in school. Only the Chinese group shows a larger variation of AoA between 3 and 22 years (mean: 13.3). Due to this lack of variance in two out of our three experimental groups, we believe that our data are not suited to answer more in-depth questions regarding age of acquisition.

IV Results

All data were analysed using linear mixed effect models in R version 3.2.2 (R Core Team, 2015), and packages lme4 version 1.1.10 (Bates et al., 2015), LMERConvenienceFunctions version 2.10 (Tremblay and Ransijn, 2015), lmerTest version 2.0.29 (Kuznetsova et al., 2015), and lsmeans version 2.20.23 (Lenth and Herve, 2015). Mixed effect models are preferable to repeated-measures ANOVAs because they can simultaneously adjust for repeated measures of both participants and items, they can handle missing data and une-qual sample sizes in the groups more appropriately, and they can easily incorporate a continuous predictor such as proficiency (Gelman and Hill, 2007).

Each component was analyzed in three steps:

1. A model with the maximally possible fixed structure was reduced by removing all non-significant higher order effects through log-likelihood ratio comparisons (LMERConvenienceFunctions::bfFixefLMER_F.fnc).

2. The resulting model was subjected to a type 3 ANOVA analysis, with denominator degrees of freedom estimated by Satterthwaite’s approximation (lmerTest::anova).

3. The highest order interaction was followed up with pairwise post hoc tests (lsmeans::lsmeans/lstrends) adjusted for multiple comparisons using Tukey’s honest significant differences.

All final models can be found in Appendix 1.

1 Sentence-final acceptability judgments

Boxplots by condition and group are shown in Figure 2. A linear mixed effect model predicting the percentage of accepted trials from group and condition while controlling for the random intercepts by participant revealed a significant group:cond interaction as the highest order effect (see Table 2 in Appendix 1). Pairwise post hoc contrasts show that within the English (Eng) group, A and B were accepted to the same degree (p = .99), but condition C was accepted less than A and B (both ps < .001), and condition D was accepted less than C (p < .001). Within the German (Ger) group the picture is the same, in that A and B were accepted to the same degree (p = .65), but condition C was accepted less than A and B (both ps < .001), and condition D was accepted less than C (p < .04). Within the Chinese (Chi) group on the other hand, there was again no difference between conditions A and B (p = 0.87), but also no difference between conditions C and D (p = 0.99), whereas A and B were always accepted more than C and D (all ps < .01). When analyzed by condition, we learn that Eng and Ger treat condition A and B each in the same way (all ps > .19), while the Chinese group accepts condition A and B each less often than the other groups (all ps < .02). Within condition C, we find that the English




and Chinese group accept this condition to the same degree (p = .82) but the German group accepts it less (both ps > .001). Within condition D, there is a marginal difference between Eng and Ger (p = .0518), but the Chinese group judged condition D as accept-able more often than both the English and German group (both ps < .001).

2 Event-related potentials

All maximal models included a 2-level factor to distinguish between subsets of condi-tions. For CPS effects reflecting prosodic phrasing, we (1) focused on correct conditions A and B only (to avoid any confounds with violation effects), but (2) collapsed across their respective boundary positions, resulting in factor bound(ary) that contrasted two levels: Boundary vs. no Boundary (B vs. noB). Thus, B(oundary) comprizes CPS1 in condition B and CPS2 in condition A (see Table 1), whereas noB(oundary) comprizes the corresponding absence of a CPS in the respective other condition. In contrast, for all analyses of ERP garden-path effects (N400 and P600s), we included the factor cond(ition), which contrasted the correct control and the violation condition (i.e. C vs. B, and D vs. A). As between-participant factors we included the three-level factor group (Eng(lish), Ger(man), Chi(nese)) as well as the continuous proficiency score (d′ between condition D and A). The proficiency score was not mean-centered because it was the only continu-ous variable included in the analysis. Please note that we explicitly refrained from dichotomizing the variable proficiency because it is naturally a continuous variable. Not only are cut-off points for ‘high’ and ‘low’ proficiency groups arbitrary and thus not comparable between studies, it has also been shown that the practice of dichotomizing a continuous variable leads to decreased statistical power, higher occurrence of false posi-tives, or an overestimation of effect sizes depending on the data set (Cohen, 1983; MacCallum, Zhang, Preacher, and Rucker, 2002).

To investigate the spatial distribution of the ERP components, all 19 electrodes (see voltage maps in Figure 3 for positions) were assigned to a three by three array of regions of interest, organized by a horizontal and a vertical axis. The horizontal axis was opera-tionalized as the factor Hem(isphere) with the three levels L(eft), M(idline) and R(ight). The vertical axis was divided up as factor AntPost with levels Ant(erior), including Fp1/2, F7/8, F3/4, and Fz, Cen(tral), including T3/4, C3/4 and Cz, and Post(erior), including T5/6, P3/4, O1/2 and Pz. Thus, the full models had the structure µV ~ cond/bound*prof*group*Hem*AntPost. When trying to account for as much by-participant and by-item variability as possible, we encountered the unfortunate situation that when even just one random slope was included, some models did not converge (i.e. the algo-rithm did not find a satisfying fit for the data). In order to keep the models comparable, we therefore removed random slopes from all models. Thus, all models described below include only random intercepts for participant and item.

Because for all components the highest order interactions in the full model always included the factor AntPost, we decided to follow up these interactions by running separate models for the Ant, Cen, and Post levels, and, for brevity, only those models are shown in this article (see Appendix 1). This practice – rather than splitting by group – enabled us to directly investigate interactions between boundary/condition group, and proficiency in the same model. Furthermore, in many of the models there are bound/cond:Hem interactions present. Those were always due to the component being stronger




over midline compared to lateral electrodes. Since the meaning of this term was the same in all models, we will not explicitly state this effect again for each model.

Figure 3. ERP boundary effects for the CPS component. (a) ERP waveforms of B and noB by group, arrows indicate point of time-locking. (b) Voltage maps displaying mean voltage of the B-noB difference wave in the 0 to 600 ms time window after onset of the boundary by group and within-group median split. (c) Least square linear trend, as estimated on basis of the raw data, of proficiency on the mean amplitude in B and noB in the time window average between 0 and 600 ms, shown separately by group and AntPost and averaged across the three levels of Hemisphere.Note. Shadow represents the standard error. The y-axis shows negative scores upwards to match the display of the ERPs. The vertical dotted line indicates the mean d prime score of 1.71.




3 Closure Positive Shift (CPS)

The response to the presence of a boundary can be seen in Figure 3, showing the event-related potentials by group (Figure 3a), as well as the influence of proficiency by group and the levels of AntPost (Figure 3c). The topographical voltage maps (in Figure 3b) display the ERP difference (Boundary minus noBoundary) and are divided by group and a within-group (not across groups) median split of proficiency. Please note that profi-ciency was entered as a continuous factor into analysis, and the median split was used for visualization purposes only. Inspection of the proficiency graph particularly suggests (1) that increasing proficiency levels resulted in larger CPS amplitudes in all groups, espe-cially at frontal electrodes, and (2) that the absolute size of the CPS seems to be influ-enced by group, with largest amplitudes in the Chinese group. Amplitude of the CPS component – collapsed across CPS1 and CPS2 – was quantified as the average amplitude in µV in a 0 to 600 ms time window using a −200 to −50 ms baseline, as the CPS can be triggered prior to the onset of the silence (Pauker et al., 2011).

Ant. At anterior electrodes, the bound:prof:group interaction just fails to reach signifi-cance when conducting the Satterthwaite adjusted ANOVA on the final model. Given that this interaction is thus likely not meaningful, we will continue to investigate the next lower order interactions, bound:group and bound:prof. Please note that those lower order interactions each average across the factors that are not part of the interaction, i.e. for the bound:group interaction the group means are predicted for the average of Hem and the mean value of proficiency, which is 1.71. Hence, this by-group analysis treats profi-ciency as if it was constant at 1.71 and will therefore compare the groups as if they all had a proficiency of 1.71. In order to visualize the meaning of this comparison we added a vertical line at 1.71 in Figure 3. Note also that a d′ of 1.71 is a low value for the English and German group, and a high value for the Chi group (see Figure 2). On the other hand this way of analysing group differences equalizes proficiency and provides an estimate for the group effects as if proficiency was the same between them. With this in mind, pairwise post hoc contrasts show that while the CPS is present as a positive estimate in all groups (Mean(B,Eng - noB,Eng) = 1.5 µV, t(81495) = 12.1, p < .001; Mean(B,Ger - noB,Ger) = 0.8 µV, t(81495) = 9.2, p < .001; Mean(B,Chi - noB,Chi) = 2.6 µV, t(81495) = 15.7, p < .001), the estimated CPS amplitudes all differ significantly between the groups (all ps < .01), in that the CPS is largest in the Chinese group (2.6 µV), followed by the English (1.5 µV) group, followed by the German group (0.8 µV), again, assuming a mean proficiency of 1.71 in all groups. The bound:prof interaction reveals that when the d′ increases by 1, the difference between B and noB becomes 0.5 µV more positive (Prof(B - noB) = 0.5 µV, t(81495) = 8.6, p < .001). This result is also averaged across Hem and group.

Cen. At central electrodes, the bound:prof:group interaction reaches significance. Spe-cifically, there is a significant influence of proficiency on the B-noB difference in the English (ProfB,Eng - noB,Eng = 0.56 µV, t(58178) = 5.5, p < .001) and the Chinese group (ProfB,Chi - noB,Chi = 0.45 µV, t(58178) = 3.8, p < .01), but not in the German group (ProfB,Ger

- noB,Ger = 0.12 µV, t(58178) = 1.5, p = .67). This influence of proficiency does not differ between the English and the Chinese group (p = 0.99).




Pos. At posterior electrodes as well, the bound:prof:group interaction is significant. Here, we find a significant influence of proficiency only in the English group (ProfB,Eng

- noB,Eng = 0.33 µV, t(81501) = 3.9, p < .01), but not in the German (ProfB,Ger - noB,Ger = -.08 µV, t(81501) = −1.2, p = .86) or Chinese group (ProfB,Chi - noB,Chi = 0.27 µV, t(81501) = 2.6, p = .09).

In sum, the presence of the CPS has been confirmed in all groups. Analysing the three AntPost levels suggests that the CPS is differently distributed between groups, in that it is more narrowly focused at Cen electrodes in the German group, whereas its overall amplitude seems to be largest in the Chinese group. There is a significant influ-ence of proficiency, most prominently at electrodes where the CPS is strongest, i.e. at frontal electrodes. Taken together, this pattern confirms that proficiency plays a central role for the magnitude of the CPS component, importantly, even in native speakers.

4 P600 garden-path effect in C

Figure 4 shows the P600 garden-path effects in condition C compared to B. A first inspection suggests that higher proficiency leads to a larger P600 in the English and German groups, while the Chinese group seems to display only a small P600 that is independent of proficiency. For analysis we quantified the P600 in C as the mean amplitude in a time window between 400 and 1300 ms after the onset of the second verb phrase (come running; see Table 1) relative to a baseline between −50 and 50 ms. It is obvious from Figure 4a that condition C and B in Eng and Chi differ in the 500 ms leading up to the target word. Selecting the baseline interval during this pre-stimulus time window results in enhanced positivities for the violation condition, likely inflat-ing the real effect. Following Pauker et al. (2011), we chose the −50 to 50 ms baseline as a more conservative approach to avoid overestimating the P600 effect amplitudes in any given group. The observation that in all groups and at all electrode sites, the two conditions do not diverge during the first 400 ms, confirms the appropriateness of this baseline selection.

Ant. The cond:prof:group interaction reveals that with increasing proficiency, the C-B difference increases significantly in the English (ProfC,Eng - B,Eng = 0.84 µV, t(40573) = 4.8, p < .001) and German group (ProfC,Ger - B,Ger = 0.38 µV, t(40573) = 2.9, p < .05), but not in the Chinese group (ProfC,Chi - B,Chi = −0.26 µV, t(40573) = −1.3, p = .77). A follow-up on the difference of this slope between groups reveals that the influence of profi-ciency is the same between the English and German group (p = .71).

Cen. The results at central electrodes are qualitatively the same as at anterior elec-trodes. Again, we find a cond:prof:group interaction which indicates that higher profi-ciency leads to a larger P600 effect in the English (ProfC,Eng - B,Eng = 0.96 µV, t(28947) = 5.3, p < .001) and German group (ProfC,Ger - B,Ger = 0.46 µV, t(28947) = 3.4, p < .01), but not in the Chinese group (ProfC,Chi - B,Chi = −0.31 µV, t(28947) = −1.5, p = .66). The sec-ond step follow-up shows that the proficiency slopes of the C-B difference do not differ between Eng and Ger (p = .65).




Pos. We find again a cond:prof:group interaction, which reveals a significant increase of the P600 effect with increasing proficiency in the English (ProfC,Eng - B,Eng = 1.14 µV, t(40574) = 7.7, p < .001) and German group (ProfC,Ger - B,Ger = 0.70 µV, t(40574) = 6.3,

Figure 4. P600 garden-path effects in condition C. (a) ERP waveforms for all three groups in condition C and B timelocked to the second verb phrase ‘come running’. (b) Voltage maps of the mean difference between C and B in the 400 to 1300 ms time window following ‘come running’ by group and within-group median split. (c) Estimated linear trend of proficiency on the P600 mean amplitude in the violation condition C and its correct control B by group and AntPost.




p < .0001), but not the Chinese group (ProfC,Chi - B,Chi = 0.06 µV, t(40574) = 0.4, p = .99). Again, this increase does not differ in magnitude between Eng and Ger (p = .51).

In sum, the P600 is influenced by both group and proficiency; while ERPs in the Chinese group are not influenced by proficiency, ERPs in Eng and Ger behave in a sta-tistically indistinguishable manner, and their P600 is larger, more broadly distributed and heavily influenced by proficiency at all levels of AntPos. Because the P600 in Chi is independent of proficiency, we can test whether there is any difference between C and B. This test will again assume a proficiency of 1.71, but because there is no influence of proficiency this is appropriate. It revealed that while there is no significant difference between C and B at Ant or Cen (both ps > .25), there is a positivity at Pos (MeanC,Chi - B,Chi = 0.77 µV, t(40575) = 3.1, p > .05). Thus, we can conclude that there is a small P600 elicited in the Chinese group, which is restricted to posterior sites.

5 N400 garden-path effect in D

Results for the D versus A comparison can be seen in Figure 5. The ERPs show the bi-phasic N400-P600 pattern over time, while the proficiency and topography maps are divided up by the two components in their respective time windows. The N400 was quantified as the mean amplitude in the 300 to 600 ms time window after the splicing point (see Table 1), using a −500 to 0 ms baseline. Inspection of these graphs show little variation in the N400 either by group or by proficiency.

Ant. At anterior electrodes we only see a cond:group and cond:prof interaction. First, the cond:group interaction reflects a lack of difference between condition D and A in the Eng-lish (Mean(D,Eng - A,Eng) = −0.20 µV, t(40737) = −1.1, p = .87) and Chinese group (Mean(D,Chi

- A,Chi) = 0.23 µV, t(40737) = 1.3, p = .81), but a relative positivity in the German group (Mean(D,Ger - A,Ger) = 1.16 µV, t(40737) = 8.8, p < .001). These estimates, again, are under the assumption that proficiency is fixed at 1.71. The cond:prof interaction on the other hand shows that – independent of group – a higher proficiency leads to a more positive differ-ence between D and A at anterior electrodes (Prof(D - A) = 0.22 µV, t(40735) = 2.6, p < .01).

Cen. Pairwise post hoc tests on the cond:prof:group interaction show that while there seem to be slightly different proficiency slopes overall, none of the comparisons of inter-est (i.e. D-A for Eng, Ger, Chi; D for Eng, Ger, Chi; A for Eng, Ger, Chi) showed a sig-nificant influence of proficiency (all ps > .05), although the effect for the D-A proficiency slope in Chi is marginal (ProfD,Chi - A,Chi = 0.50 µV, t(29056) = 2.7, p = .08). In light of the unclear meaning of this interaction, we continued to follow up the next lower order inter-action, i.e. cond:group. It shows that while there is no significant difference between D and A in the English (Mean(D,Eng - A,Eng) = −0.54, t(29048) = −2.75, p = .07, though marginal) and the Chinese group (Mean(D,Chi - A,Chi) = 0.00 µV, t(29048) = 0.0, p = 1), the positivity is still present in Ger (Mean(D,Ger - A,Ger) = 0.68 µV, t(29048) = 4.7, p < .001).

Pos. Like at Cen, we find a cond:prof:group interaction which can be ascribed to some difference in the proficiency slopes between conditions and groups, although, again, none of the comparisons of interest reach significance (all ps > .10). Therefore we will again turn towards the next significant lower order interaction, which is cond:group. Here, we see a significant negative difference between D and A in all groups (Mean(D,Eng




- A,Eng) = −1.3 µV, t(40720) = −8.7, p < .0001; Mean(D,Ger - A,Ger) = −0.9 µV, t(40720) = −7.4, p < .0001; Mean(D,Chi - A,Chi) = −0.86 µV, t(40720) = −3.8, p < .001), again predicted for mean proficiency. Further follow-up comparisons show that the D-A difference is the same in all groups (all ps < .73).

In sum, a significant N400 was found in all groups at posterior electrodes only, and the N400 amplitude did not differ between groups or proficiency. Furthermore, at Ant and Cen electrodes we saw a positivity in the German group.

Figure 5. N400 and P600 garden-path effects in condition D. (a) ERP waveforms, time-locked to the splicing point on the fricative “th” (see Table 1), for all three groups in condition D and its lexically matching control A. (b) Voltage maps of the mean difference between D and A, showing the N400 in the 300 to 600 ms time window, and the P600 in the 800 to 1300 ms time window. The N400 maps are not divided by proficiency, as it had no statistical influence. (c) Estimated linear trend of proficiency on the N400 and P600 mean amplitude in the violation condition D and its correct control A by group and AntPost.




6 P600 garden-path effect in D

The second component elicited as a response to the superfluous boundary in D was a P600. It was quantified as the mean amplitude in an 800 to 1300 ms time window time-locked to the splicing point, again using a −500 to 0 baseline. A look at Figure 5 suggests that higher proficiency leads to a larger P600 amplitude in all groups, and that the P600 is in general larger in the English and German group.

Ant. The highest order effect involving condition is an interaction with group, which revealed a significant, positive D-A difference only for the English (Mean(D,Eng - A,Eng) = 0.7 µV, t(40728) = 3.3, p < .05) and the German group (Mean(D,Ger - A,Ger) = 1.8 µV, t(40728) = 11.1, p < .001), but not for the Chinese group (Mean(D,Ger - A,Ger) = –0.6 µV, t(40728) = −2.6, p = .10). Follow-ups on this difference revealed that the difference between D and A was smaller in the English than in the German group (Mean(D,Eng - A,Eng)

- (D,Ger - A,Ger) = −1.1 µV, t(40728) = −4.4, p > .01). There was no significant influence of proficiency at the Ant level.

Cen. Here we find for the first time a cond:prof:Hem interaction which indicates that proficiency has a different influence on condition depending on the level of hemisphere. Specifically, proficiency has no influence on the D-A slope in the left hemisphere (ProfD,L

- A,L = 0.32 µV, t(29033) = 2.3, p = .19) while it does at the midline (ProfD,M - A,M = 1.2 µV, t(29033) = 6.4, p < .0001) and in the right hemisphere (ProfD,R - A,R = 0.41 µV, t(29033) = 3.0, p > .05). There is further a cond:prof:group interaction reflecting the fact that profi-ciency only influenced the condition difference in the English (ProfD,Eng - AEng = 0.64 µV, t(29050) = 3.5, p < .01) and the Chinese group (ProfD,Chi - A,Chi = 1.0 µV, t(29050) = 4.8, p < .001), but not the German group (ProfD,Ger - A,Ger = 0.28 µV, t(29050) = 1.9, p = .38). Further, the influence of proficiency is the same between Eng and Chi (p = .99).

Pos. Here we find again a cond:prof:group interaction which was caused by proficiency only having an impact on the D-A difference in the English (ProfD,Eng - A,Eng = 0.89 µV, t(40724) = 5.6, p < .001) and Chinese group (ProfD,Chi - A,Chi = 1.3 µV, t(40724) = 7.5, p < .001), but not in the German group (ProfA,Ger - D,Ger = 0.26 µV, t(40724) = 2.2, p = .25), and again this effect is the same in Eng and Chi (p = .84).

In sum, the P600 is present in all groups, but it is larger and more broadly distributed in Eng and Ger. In Chi, where the effect is smallest, there is even an inversion of polarity at low levels of proficiency, in that there is a relative negativity between in D in compari-son to A (see Figure 5).

V Discussion

1 Boundary processing: the CPS

All groups elicited the CPS component reflecting their successful processing of pro-sodic boundaries, although the strength of the effect was influenced by factor group as well as proficiency level. The Chinese group showed overall the largest amplitude,




followed by the English and then the German group. These group differences were not necessarily expected given that all three investigated languages use intonational phrases. The larger CPS in the Chinese group might be related to the fact that they show the low-est proficiency overall, and may have had substantial difficulties both with the semantic contents and syntactic structures of the presented sentences (partly due to phonological differences between English and Chinese, such as a different phoneme inventory and the presence of lexical tones in Chinese). Thus, it might be that the Chinese group focused more on aspects that they could process easily, namely prosody, in a similar manner in which children with limited semantic and syntactic knowledge use prosody to discover syntactic boundaries during prosodic bootstrapping (e.g. Morgan and Demuth, 2014). However, there is no evidence for enhanced CPS amplitudes in young children, and the numeric amplitude difference between Chinese and English speakers as a group is minimal (see Figure 3). The relatively large CPS in Chinese learners may, therefore, be primarily taken as support for the notion that boundary processing as such can be done at rather low levels of language proficiency, most likely because the under-lying neurocognitive mechanisms are shared with other cognitive domains such as music (Glushko et al., 2016; Steinhauer et al., 2009). The finding of a smaller CPS in the German group was surprising and contrasts with Nickels et al. (2013), a study that used the same paradigm but only tested a group of highly proficient German speakers against the native speaker group. There, the CPS components were undistinguishable between groups, suggesting that at high levels of proficiency the CPS component is the same. It appears that the inclusion of overall lower proficiency participants in the cur-rent study led to the reduced CPS amplitude within the German group. However, since even the less proficient Chinese group displayed a large CPS, this interpretation must be viewed with caution.

At anterior electrodes we also find that higher proficiency led to a larger CPS compo-nent in the B-noB difference wave in all groups. It should be noted, however, that this ERP pattern is somewhat ambiguous as it may be driven by both conditions: Figure 3 shows that increased CPS amplitudes are associated with both a more positive boundary condition, as well as a more negative noB condition. Previous studies have found that CPS components are usually preceded by negativities, which in the case of the CPS2 at the second boundary amounts to a substantial, ramp-like negativity at frontal electrodes (Nickels et al., 2013; Pauker et al., 2011). This negativity is associated with the expec-tancy of an upcoming boundary. In our case we averaged across both CPS components, however, it is likely that the increased CPS amplitude with increasing proficiency is caused by both a stronger expectancy negativity in noB (for CPS2), as well as a larger positive shift in the boundary position.

2 Garden path effects and P600 in condition C

All groups accepted garden path condition C (without any boundary) to a lesser extent than its correct matched control condition B, whose early prosodic boundary supported the required early closure (EC) interpretation. Interestingly, condition C was more often rejected as ‘unnatural’ by German participants than by English native speakers. This pat-tern is exactly what one would expect if Germans were unable to process tense and




aspect of the progressive form of the first verb (was approaching) and processed it instead like the past tense form (approached), which supports the (inappropriate) late closure interpretation (Frazier et al., 2006). However, when an early prosodic boundary is present (in condition B), it overrides any initial preferences in favor of EC (Itzhak et al., 2010) and the group differences largely disappear (and do not reach significance anymore). Overall, these data provide an interesting and novel case of asymmetric inter-actions between prosodic and morphosyntactic processing in L2 learners. Unlike the German group, Chinese L2 learners did not reject C more than English speakers; how-ever, their ability to discriminate between B and C was very limited in general, suggest-ing that issues related to tense or aspect played a minor role compared to other morphosyntactic challenges. This interpretation is supported by the ERP online data.

We found a significant P600 effect elicited at the lexically disambiguating second verb phrase come running in all groups (for a discussion about lexical vs. prosodic dis-ambiguation see (Bögels et al., 2013), but the P600 in the Chinese group was very small and not modulated by proficiency (potentially because of the rather small size). In the other two groups the P600 was large and increased in amplitude with increasing profi-ciency. This latter relationship mirrors previous findings in other ERP studies on L2 morphosyntactic processing (e.g. Steinhauer et al., 2009) and supports the notion that P600 amplitudes partly reflect the degree of grammaticalized processing of a given structure. Most previous studies have reported this relationship for L2 learners, but more recent work has demonstrated that ERPs may reflect varying levels of proficiency even in native speakers (e.g. Pakulak and Neville, 2010), in line with our present data. The finding that higher proficiency resulted in larger P600s in the German group may be viewed as counter-intuitive: if German learners perceived condition C as a stronger gar-den-path than native speakers because they were unable to process the aspect of the progressive verb form in a native-like manner (see above), then more native-like profi-ciency should actually weaken the garden-path effect and reduce their P600 amplitudes. It is possible that even higher levels of proficiency are required in order to overcome the problems German speakers of English have with aspect. In other words, the fact that Germans elicited a P600 of similar magnitude as native speakers (despite their overall lower level of proficiency) is likely due to Germans’ problems with English aspect (i.e. their stronger preference for late closure), but the P600 increase as a function of L2 pro-ficiency is likely driven by more native-like mechanisms when revising the structure. If correct, this interpretation would lend support to current ideas according to which P600s are comprized of subcomponents that reflect various cognitive operations (e.g. Friederici, Mecklinger, Spencer, Steinhauer, and Donchin, 2001; Steinhauer and Connolly, 2008). These operations may include diagnosis and reanalysis of syntactic anomalies (Friederici et al., 2001), the integration of multiple information streams (Brouwer and Hoeks, 2013; Kuperberg, Kreher, Sitnikova, Caplan, and Holcomb, 2007; van Herten, Chwilla, and Kolk, 2006), and task-related evaluations (Sassenhagen et al., 2014).

Regarding the significant relationship between proficiency and P600, it is worth not-ing that we obtained the results here for an ERP effect that was calculated between condi-tions C and B, while we calculated the proficiency score as the d prime between D and A. This finding might thus lend more validity to the proficiency score being a good operationalization for proficiency, and not a relationship between the ERP effects and the end-of-sentence judgments in the same conditions. The lack of proficiency effect in the




Chinese group might be related to their level of proficiency generally being low, and during this relatively mild garden path (especially in comparison to D) only at higher levels one would see a stronger divergence between the conditions.

3 Garden-path effects in condition D: N400 and P600

In line with previous studies (Nickels et al., 2013; Pauker et al., 2011) and with the boundary deletion hypothesis (Pauker et al., 2011), condition D with its superfluous early boundary in a late closure sentence was rated as the least natural condition by English and German listeners, and was rejected more often than condition C. In contrast, Chinese participants rejected this condition less often; here the ratings were comparable to those in condition C. This latter finding suggests that, contrary to the boundary dele-tion hypothesis, Chinese listeners did not perceive a superfluous boundary as more dif-ficult to process than a missing boundary (in condition C). ERP patterns shed light on the underlying online processes.

N400. All three groups showed a significant N400 effect in condition D. Interestingly, this was the only component that was not influenced by either group or proficiency. The N400 component is elicited by the prosodically detached noun phrase which cannot be integrated into either the previous or the following phrase. As explained by Pauker et al. (2011), this mismatch between syntax and prosody likely prevented the assignment of a theta role to the detached noun. Due to the severity of the violation it is conceivable that the garden path effect was equally strong for all participants, independent of language background and proficiency. As in previous studies, the N400 was most prominent at posterior electrodes. In the German group, frontal electrodes already displayed an early positivity, most likely the first indication of the subsequent P600 effect.

P600. The P600 in D was the second response to the severe, prosodically induced garden-path. It was found to be generally larger and more broadly distributed in the English and German groups. Also, a significant influence of proficiency on the P600 dif-ference wave was only found for English and Chinese groups. However, to say that proficiency had no influence in the German group is not correct; instead, higher levels of proficiency shifted the ERPs in both conditions to the more positive amplitude range. While the effect of proficiency on condition D is the same in all groups, only the German group shows a similar shift in condition A as well. This is important because the positiv-ity observed in D is arguably the sum of two effects: (a) the CPS elicited at the second boundary (which should in principle be of the same amplitude as the one elicited in A), and (b) the P600 proper elicited by the prosody-syntax mismatch (for details, see Pauker et al., 2011). This fact is relevant because the CPS analysis revealed that the CPS is of different size in the different groups, and in fact we see in Figure 5a that the positivity (CPS) in control condition A is largest in the Chinese group, though, this effect is strong-est at frontal electrodes. We also know from an analysis not shown in this paper that in the German group this influence of proficiency is particularly strong for the CPS2, which explains the strong increase for the German group in condition A. The consequence then of calculating the P600 in D as the difference wave between D and A is that for both the Chinese and German group the additional amplitude increase of the P600 proper in D




would need to be proportionally larger in order to offset the effect of the larger CPS. In this respect, the P600 group effect found in the Chinese group might be an underestima-tion of the real effect, and so might be the lack of the influence of proficiency in the German group. Moreover, the German proficiency effect is numerically positive at Cen and Pos at ca. 25 µV, though this is not large enough to reach significance. Therefore we refrain from interpreting this effect as strong evidence for a total lack of proficiency influence in the Germany group.

It is curious that while the N400 amplitude is the same size in all groups, implying that the retrieval and integration of the floating NP is equally difficult for all groups, the P600 amplitude does differ between the groups. Why is the P600 effect smaller in the Chinese group when the N400 is the same? One possibility is that in the Chinese group, the N400 reflects less the failure to assign a thematic role to the NP, and more problems with the lexical retrieval of the NP, which would be in line with the generally lower proficiency in this group. An alternative account could be that Chinese learners were able to identify the detached noun phrase (thus eliciting an N400), but were not able to compute a struc-tural revision that could solve the problem (thus no P600). The absence of P600s for morphosyntactic violations – sometimes in presence of N400s – has been found in a number of studies on low proficiency L2 learners (e.g. Osterhout, McLaughlin, Pitkänen, Frenck-Mestre, and Molinaro, 2006; Steinhauer et al., 2009). The strikingly high accept-ability ratings for condition D in the Chinese group would be compatible with this inter-pretation. From this perspective, the apparent incompatibility of this group’s data with the boundary deletion hypothesis may be somewhat misleading. That is, had Chinese participants actually attempted to delete a superfluous prosodic boundary (in order to revise the sentence structure), ERPs and behavioral data may have confirmed that this process is more effortful than mentally creating a boundary. However, additional data from more proficient Chinese-English bilinguals would be needed to answer this ques-tion. In line with this view, we see that in the Chinese group the garden path effect devel-ops from a negativity into a positivity at parietal electrodes with increasing proficiency (see Panel Post, Chi in Figure 5c). This might be a perseverance of the previous N400 in the lower proficient Chinese participants, which is superimposed by a P600 only in the higher proficiency speakers.

In general, it is worth noting that the Chinese group was tested while immersed in their L2 (i.e. they were living in Canada or the UK), while the German group was not (they were living in Germany). Interestingly, the average proficiency was nevertheless overall higher in the German group, despite limited immersion. On the one hand, this finding is reminiscent of similar recent reports on ‘native-like’ patterns in late L2 learn-ers that were never subjected to L2 immersion (e.g. Bowden, Steinhauer, Sanz, and Ullman, 2013). As in the present case, such findings seem to be limited to relatively sali-ent violations and L1-L2 pairings that allow for transfer from L1. On the other hand, the fact that even after a long time of immersion in Canada the Chinese group did not reach similar proficiency levels as the English or German groups speaks to the limiting effects of very different L1 backgrounds.

Moreover, we have demonstrated the merits of using linear mixed effect models to analyze a data set such as ours, which we believe to be representative of many second language studies. The linear mixed effect model makes adjustments to compensate for the widely unbalanced number of participants in each group, which is much more




difficult to deal with using traditional ANOVAs, and is likely to violate the homogeneity of variance assumption. It further adjusts for random intercepts of items and slopes, although unfortunately in our case a fuller random structure was not supported by the data. And lastly, mixed models allowed us to organically include a continuous variable such as proficiency. This becomes particularly relevant in cases where group and profi-ciency is confounded, as it happens frequently in second language research. For a tradi-tional ANOVA, the groups would have to be divided into high and low proficiency groups depending on an across group median split, which, in our case, would have resulted in very few Chinese speakers in the high proficiency ‘group’ and very few German and English speakers in the low proficiency ‘group’. Again, this would have likely led to heterogeneous variance between the groups. Such an analysis would only allow additional conclusions about the difference between two groups, but not about the proportional increase of the effect with increasing proficiency. Lastly, the inclusion of a continuous predictor allows the results to be easily generalized to other samples once their proficiency values are known. However, assuming the influence of a predictor vari-able to be linear might be an oversimplification in itself. In such a case, more sophisti-cated models (e.g. Generalized Additive Mixed Models; see Wood, 2006) might be necessary to accurately describe the data. Furthermore, the strong effects of proficiency found in the native speakers in all components except the N400 show how misleading it would be to assume (or set) native speakers to have a perfect performance/proficiency.

While it is clear that phonology and prosody are often ignored in the classroom, our data show that prosodic information, especially in combination with certain morphosyn-tactic processing patterns, can play an important role in L2 learners’ command of their second language. For example, participants in our German group would likely have prof-ited from some instruction addressing the relationship between English aspect, prosodic phrasing, and attachment preferences. More recently, there have been first attempts to work towards including specific prosody-related exercises into language acquisition curricula. For example, Trouvain and Gut (2007) recommend the following exercises: ‘(1) introduction of the topic, e.g. with a comprehension text, (2) listening control, i.e. dif-ferentiate (compare) and identify (recognize) prosodic features, (3) imitation attempts, indi-vidually and in chorus in order to rehearse anonymously, (4) correction of deviant forms, to make the learners aware of the critical phonetic features, (5) repeated listening control, (6) further imitation attempts with feedback, (7) automation by repeating, reading, vari-ation of speaking style’ (p. 179 ff). They further point out the importance of tailoring the exercises to the L1 background of the learners, in pointing out that native speakers of tonal languages typically have greater difficulties in acquiring the phonology of a Germanic language. Our data suggest that the list of exercises should be complemented by exercises focusing on the integration of prosodic and structural information. This includes practic-ing the appropriate placement and realization of prosodic boundaries (e.g. in terms of strength and acoustic features) in speech production as well as exercises that help high-light the complex interplay of prosody, syntax and aspect in speech perception.

Acknowledgements

We would like to thank Elizabeth Murphy, Catherine Royea and Jessica Cooper for their valuable help with data acquisition. We are also deeply thankful to Martijn Wieling for his insightful advice on details about random structure definition and issues of multicollinearity. We also owe special




thanks to Russell V. Lenth for his kind help on details on the post hoc comparisons used on the final models.

Declaration of conflicting interest

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: 1. Canada Research Chair program/CIHR (Project # 950-209843) - PI: Karsten Steinhauer; 2. Canada Foundation for Innovation (CFI/CRC project # 201876 - PI: Karsten Steinhauer); 3. NSERC (Grant # RGPIN 402678-11 - PI: Karsten Steinhauer, Project on “Brain signatures of nativeness in second language acquisition II”); 4. FQRSC Team Grant (Grant # 2016-SE-188196 - PI: Lydia White, co-PI: Karsten Steinhauer, Project on “Perspectives neurocognitives sur l’acquisition, la perte et le traitement du langage” (2016-SE-188196)).

Notes

1. Problems with progressive aspect are particularly notorious. For example, the use of adverbs and other constructions in German to express ongoing actions (e.g. er war gerade dabei, Kaffee zu machen ‘he was in the process of making coffee’ for ‘he was making coffee’) is not mandatory: often progressive aspect is not expressed at all in German, and the past tense (er machte Kaffee) is used to refer to both English simple past tense and progressive. When speaking English, German learners often replace the progressive form with simple past tense (‘he made coffee’ instead of ‘he was making coffee’), and in perception the difference between the two forms often remains unclear to them (Dürich, 2005).

2. The two cited studies used the same English materials as the present investigation. However, there is also convergent evidence for stronger problems with superfluous compared to missing boundaries from (1) an auditory study in Dutch (Bögels et al., 2013) and from reading studies in German (Steinhauer, 2003), but see also Zahn and Scheepers (2015) for similar findings.

References

Alderson JC (1980) Native and nonnative speaker performance on cloze tests. Language Learning 30: 59–76.

Bates D, Maechler M, Bolker B, et al. (2015) lme4: Linear mixed-effects models using ‘Eigen’ and S4: Version 1.1–9. Available at: https://cran.r-project.org/web/packages/lme4/index.html (accessed May 2016).

Bley-Vroman R (1989) What is the logical problem of foreign language learning? In: Bley-Vroman R (ed.) Linguistic perspectives on second language acquisition. Cambridge: Cambridge University Press, pp. 41–68.

Bögels S, Schriefers H, Vonk W, and Chwilla DJ (2011) Prosodic breaks in sentence processing investigated by event-related potentials. Language and Linguistics Compass 5: 424–40.

Bögels S, Schriefers H, Vonk W, Chwilla DJ, and Kerkhofs R (2013) Processing conse-quences of superfluous and missing prosodic breaks in auditory sentence comprehension. Neuropsychologia 51: 2715–28.

Bowden HW, Steinhauer K, Sanz C, and Ullman MT (2013) Native-like brain processing of syntax can be attained by university foreign language learners. Neuropsychologia 51: 2492–2511.


https://cran.r-project.org/web/packages/lme4/index.html



Brouwer H and Hoeks JCJ (2013) A time and place for language comprehension: Mapping the N400 and the P600 to a minimal cortical network. Frontiers in Human Neuroscience 7(758): 1–12.

Cohen J (1983) The cost of dichotomization. Applied Psychological Measurement 7: 249–53.Delorme A and Makeig S (2004) EEGLAB: An open source toolbox for analysis of single-trial

EEG dynamics including independent component analysis. Journal of Neuroscience Methods 134: 9–21.

Dürich K (2005) The acquisition of the English tense and aspect system by German adult learners. Unpublished Master’s thesis, Technische Universität Chemnitz, Chemnitz, Germany.

Frazier L (1987) Sentence processing: A tutorial review. In: Coltheart M (ed.) Attention and per-formance 12: The psychology of reading. Hillsdale, NJ: Lawrence Erlbaum, pp. 559–86.

Frazier L, Carminati MN, Cook AE, Majewski H, and Rayner K (2006) Semantic evaluation of syntactic structure: Evidence from eye movements. Cognition 99: B53–B62.

Friederici AD, Mecklinger A, Spencer KM, Steinhauer K, and Donchin E (2001) Syntactic pars-ing preferences and their on-line revisions: A spatio-temporal analysis of event-related brain potentials. Cognitive Brain Research 11: 305–23.

Gelman A and Hill J (2007) Data analysis using regression and multilevel/hierarchical models: Volume 1. Cambridge: Cambridge University Press. Available at: http://lac.essex.ac.uk/local-files/meetings1213/gelman_1.pdf (accessed May 2016).

Glushko A, Steinhauer K, DePriest J, and Koelsch S (2016) Neurophysiological correlates of smu-sical and prosodic phrasing: Shared processing mechanisms and effects of musical expertise. PLoS ONE 11(5): e0155300.

Hwang H and Schafer AJ (2006) Prosodic effects in parsing early vs. late closure sentences by sec-ond language learners and native speakers. Proceedings of the 3rd International Conference on Speech Prosody (pp. 585–88). Available at: http://www2.hawaii.edu/~aschafer/Hwang&Schafer2006_SP.pdf (accessed May 2016).

Itzhak I, Pauker E, Drury JE, Baum SR, and Steinhauer K (2010) Event-related potentials show online influence of lexical biases on prosodic processing. NeuroReport 21: 8–13.

Kjelgaard MM and Speer SR (1999) Prosodic facilitation and interference in the resolution of tem-porary syntactic closure ambiguity. Journal of Memory and Language 40: 153–94.

Kuperberg GR, Kreher DA, Sitnikova T, Caplan DN, and Holcomb PJ (2007) The role of animacy and thematic relationships in processing active English sentences: Evidence from event-related potentials. Brain and Language 100: 223–37.

Kuznetsova A, Brockhoff PB, and Christensen RHB (2015) lmerTest: Tests in linear mixed effects models: Version 2.0–29. Available at: https://cran.r-project.org/web/packages/lmerTest/index.html (accessed May 2016).

Lenneberg EH (1967) Biological foundations of language. London: Wiley.Lenth R and Herve M (2015) lsmeans: Least-squares means: Version 2.20–2. Available at: https://

cran.r-project.org/web/packages/lsmeans/index.html (accessed May 2016).Lopez-Calderon J and Luck SJ (2014) ERPLAB: An open-source toolbox for the analysis of event-

related potentials. Frontiers in Human Neuroscience 8: 213.MacCallum RC, Zhang S, Preacher KJ, and Rucker DD (2002) On the practice of dichotomization

of quantitative variables. Psychological Methods 7: 19–40.Mennen I (2007) Phonological and phonetic influences in non-native intonation. In: Trouvain J

and Gut U (eds) Trends in linguistics studies and monographs: Volume 53. Berlin and New York: Mouton de Gruyter, pp. 53–76.

Morgan JL and Demuth K (2014) Signal to syntax: Bootstrapping from speech to grammar in early acquisition. New York and London: Psychology Press.


http://lac.essex.ac.uk/local-files/meetings1213/gelman_1.pdf

http://lac.essex.ac.uk/local-files/meetings1213/gelman_1.pdf

http://www2.hawaii.edu/~aschafer/Hwang&Schafer2006_SP.pdf

http://www2.hawaii.edu/~aschafer/Hwang&Schafer2006_SP.pdf

https://cran.r-project.org/web/packages/lmerTest/index.html

https://cran.r-project.org/web/packages/lmerTest/index.html

https://cran.r-project.org/web/packages/lsmeans/index.html

https://cran.r-project.org/web/packages/lsmeans/index.html



Nickels S, Opitz B, and Steinhauer K (2013) ERPs show that classroom-instructed late second language learners rely on the same prosodic cues in syntactic parsing as native speakers. Neuroscience Letters 557: 107–11.

Oldfield RC (1971) The assessment and analysis of handedness: The Edinburgh inventory. Neuropsychologia 9: 97–113.

Osterhout L, McLaughlin J, Pitkänen I, Frenck-Mestre C, and Molinaro N (2006) Novice learners, longitudinal designs, and event-related potentials: A means for exploring the neurocognition of second language processing. Language Learning 56: 199–230.

Pakulak E and Neville HJ (2010) Proficiency differences in syntactic processing of monolingual native speakers indexed by event-related potentials. Journal of Cognitive Neuroscience 22: 2728–44.

Pauker E, Itzhak I, Baum SR, and Steinhauer K (2011) Effects of cooperating and conflicting prosody in spoken English garden path sentences: ERP evidence for the boundary deletion hypothesis. Journal of Cognitive Neuroscience 23: 2731–51.

R Core Team (2015) R: A language and environment for statistical computing. Vienna: R Foundation for Statistical Computing. Available at: https://www.r-project.org (accessed May 2016).

Saito K (2012) Effects of instruction on L2 pronunciation development: A synthesis of 15 quasi-experimental intervention studies. TESOL Quarterly 46: 842–54.

Sassenhagen J, Schlesewsky M, and Bornkessel-Schlesewsky I (2014) The P600-as-P3 hypoth-esis revisited: Single-trial analyses reveal that the late EEG positivity following linguistically deviant material is reaction time aligned. Brain and Language 137: 29–39.

Schafer AJ, Speer SR, Warren P, and White SD (2000) Intonational disambiguation in sentence production and comprehension. Journal of Psycholinguistic Research 29: 169–82.

Shi M (2011) A corpus based contrastive approach for the analysis of tense and aspect in transla-tion from English into Mandarin Chinese. Unpublished doctoral dissertation, Westfälische Wilhelms-Universität Münster, Münster, Germany.

Steinhauer K (2003) Electrophysiological correlates of prosody and punctuation. Brain and Language 86: 142–64.

Steinhauer K (2014) Event-related potentials (ERPs) in second language research: A brief intro-duction to the technique, a selected review, and an invitation to reconsider critical periods in L2. Applied Linguistics 35: 393–417.

Steinhauer K and Connolly JF (2008) Event-related potentials in the study of language. In: Stemmer B and Whitaker H (eds) Handbook of the cognitive neuroscience of language. New York: Elsevier, pp. 91–104.

Steinhauer K, Alter K, and Friederici AD (1999) Brain potentials indicate immediate use of pro-sodic cues in natural speech processing. Nature Neuroscience 2: 191–96.

Steinhauer K, White EJ, and Drury JE (2009) Temporal dynamics of late second language acquisition: Evidence from event-related brain potentials. Second Language Research 25: 13–41.

Tremblay A and Ransijn J (2015) LMERConvenienceFunctions: Model selection and post-hoc analysis for (G)LMER models: Version 2.10. Available at: https://cran.r-project.org/web/packages/LMERConvenienceFunctions/index.html (accessed May 2016).

Trouvain J and Gut U (2007) Non-native prosody: Phonetic description and teaching practice. Berlin and New York: Walter de Gruyter.

van Herten M, Chwilla DJ, and Kolk HHJ (2006) When heuristics clash with parsing routines: ERP evidence for conflict monitoring in sentence perception. Journal of Cognitive Neuroscience 18: 1181–97.


https://www.r-project.org

https://cran.r-project.org/web/packages/LMERConvenienceFunctions/index.html

https://cran.r-project.org/web/packages/LMERConvenienceFunctions/index.html



Weber-Fox CM and Neville HJ (1996) Maturational constraints on functional specializations for language processing: ERP and behavioral evidence in bilingual speakers. Journal of Cognitive Neuroscience 8: 231–56.

Wood S (2006) Generalized additive models: An introduction with R. Boca Raton, FL: CRC Press.Wurm LH and Fisicaro SA (2014) What residualizing predictors in regression analyses does (and

what it does not do). Journal of Memory and Language 72: 37–48.Zahn D and Scheepers C (2015) Overt prosody and plausibility as cues to relative-clause attach-

ment in English spoken sentences. PeerJ PrePrints 3: e148.

Appendix 1

Linear mixed effect models

Shown below are the final models for the behavioral, end of sentence judgments (Table 2) and the four final models calculated per component, one for each level of AntPost (Tables 3–6), subjected to a type III ANOVA (for details, see Section III.5).

Table 2. ANOVA result on the final linear mixed effect model for percentage of accepted trials.

SumSq MeanSq NumDF DenDF F Pr(>F)

group 5734 2867 2 86 9.618 0.0001699***cond 147045 49015 3 258 164.439 < 2.2e–16***group:cond 35238 5873 6 258 19.703 < 2.2e–16***

Significant terms are marked as follows: ‘.’: 0.10 to 0.05; ‘*’, 0.05 to 0.01; ‘**’: 0.01 to 0.001; ‘***’: 0.001.

Table 3. CPS: ANOVA result on the final linear mixed effect models for the mean amplitude between 0 and 600 ms in B and noB.


Ant:bound 3133.5 3133.5 1 81493 54.360 1.69E–13***prof 31.9 15.9 1 95 0.276 0.7590475group 913.5 456.8 2 83 7.924 0.0007084***Hem 442.5 221.2 2 81486 3.838 0.0215418*bound:prof 3358.5 3358.5 1 81499 58.263 2.32E–14***bound:group 4202.1 2101.1 2 81495 36.449 2.22E–16***prof:group 181.5 90.8 2 89 1.575 0.2128146bound:Hem 1117.5 558.7 2 81486 9.693 6.18E–05***prof:Hem 123.3 123.3 2 81486 2.140 0.1435212bound:prof:group 204.2 204.2 2 81500 3.543 0.0597962.Cen:bound 5160.5 5160.5 1 58173 113.695 < 2.2e–16***prof 63.7 63.7 1 84 1.403 0.239547group 373.2 186.6 2 83 4.111 0.019859*

(Continued)





Hem 358.9 179.4 2 58165 3.954 0.019191*bound:prof 1866.7 1866.7 1 58177 41.126 1.44E–10***bound:group 261.8 130.9 2 58175 2.884 0.055945.prof:group 27.7 13.8 2 83 0.305 0.738014bound:Hem 1054.4 527.2 2 58165 11.615 9.05E–06***prof:Hem 304.8 152.4 2 58165 3.357 0.034837*bound:prof:group 605.2 302.6 2 58178 6.666 0.001274**Pos:bound 4440.9 4440.9 1 81496 100.951 < 2.2e–16***prof 40.8 40.8 1 82 0.928 0.3382855group 127.5 63.8 2 82 1.449 0.2406765Hem 1040.7 520.4 2 81488 11.829 7.31E–06***bound:prof 553.2 553.2 1 81501 12.575 0.0003911***bound:group 37.6 18.8 2 81498 0.428 0.6520178prof:group 63.6 31.8 2 82 0.723 0.4883782bound:Hem 1534.5 767.2 2 81488 17.441 2.67E–08***bound:prof:group 753.2 376.6 2 81501 8.561 0.0001916***

Table 3. (Continued)

Table 4. P600 in C: ANOVA result on the final linear mixed effect models for the mean amplitude between 400 and 1300 ms in C and B.


Ant:cond 27.42 27.42 1 40572.0 0.3186 5.72E–01prof 28.15 28.15 1 84.0 0.3272 0.568851group 48.35 24.18 2 83.0 0.2810 0.755758Hem 2207.64 1103.82 2 40560.0 12.8281 2.70E–06***cond:prof 899.97 899.97 1 40571.0 10.4590 1.22E–03**cond:group 2703.41 1351.71 2 40574.0 15.7089 1.52E–07***prof:group 429.77 214.88 2 83.0 2.4973 0.088476.prof:Hem 1215.40 607.70 2 40560.0 7.0624 8.58E–04***cond:prof:group 1508.19 754.10 2 40573.0 8.7637 0.000157***Cen:cond 58.40 58.39 1 28945.3 0.8857 0.346652prof 83.80 83.76 1 82.8 1.2704 0.262952group 46.00 23.02 2 82.9 0.3491 0.706321Hem 3957.80 1978.89 2 28931.9 30.0151 9.50E–14***cond:prof 878.20 878.22 1 28945.0 13.3205 2.63E–04***cond:group 1426.80 713.41 2 28947.7 10.8207 2.01E–05***prof:group 239.80 119.89 2 82.7 1.8185 0.168694cond:Hem 640.60 320.32 2 28931.9 4.8585 7.77E–03**cond:prof:group 1425.60 712.79 2 28946.7 10.8114 2.03E–05***




Table 5. N400 in D: ANOVA result on the final linear mixed effect models for the mean amplitude between 300 and 600 ms in D and A.


Ant:cond 1.50 1.49 1 40728.0 0.0202 8.87E–01prof 38.90 38.88 1 87.0 0.5293 0.4688401group 1501.20 750.60 2 85.0 10.2201 0.0001054***Hem 52.10 26.03 2 40709.0 0.3545 0.7015497cond:prof 493.40 493.38 1 40735.0 6.7178 9.55E–03**cond:group 3235.50 1617.75 2 40737.0 22.0270 2.75E–10***prof:Hem 492.30 246.14 2 40709.0 3.3514 0.0350433*Cen:cond 19.14 19.14 1 29047.6 0.3481 0.5552181prof 33.13 33.13 1 83.3 0.6025 0.4398149group 814.29 407.15 2 84.0 7.4054 0.0010915**Hem 960.31 480.16 2 29033.5 8.7334 1.62E–04***cond:prof 52.13 52.13 1 29059.4 0.9482 0.3301889cond:group 1544.39 772.20 2 29048.3 14.0452 8.00E–07***prof:group 138.74 69.37 2 83.2 1.2617 2.89E–01cond:Hem 420.22 210.11 2 29033.5 3.8216 0.0219045*group:Hem 626.84 156.71 4 29033.5 2.8503 0.0224228*cond:prof:group 547.79 273.89 2 29056.0 4.9818 0.0068678**Pos:cond 2476.48 2476.48 1 40719.0 46.6920 8.42E–12***prof 0.40 0.40 1 83.0 0.0070 0.931294group 587.75 293.88 2 84.0 5.5410 5.49E–03**Hem 1681.48 840.74 2 40705.0 15.8520 1.31E–07***cond:prof 0.75 0.75 1 40730.0 0.0140 0.905186cond:group 614.69 307.34 2 40720.0 5.7950 0.003046**prof:group 162.92 81.46 2 83.0 1.5360 2.21E–01group:Hem 629.86 157.46 4 40705.0 2.9690 0.018312*cond:prof:group 555.52 277.76 2 40727.0 5.2370 0.00532**


Pos:cond 242.00 242.00 1 40572.0 3.8270 0.05043.prof 140.10 140.10 1 83.0 2.2150 0.14044group 52.20 26.10 2 83.0 0.4130 0.66296Hem 1801.90 901.00 2 40559.0 14.2510 6.50E–07***cond:prof 3648.50 3648.50 1 40572.0 57.7110 3.11E–14***cond:group 1710.10 855.00 2 40575.0 13.5250 1.34E–06***prof:group 43.70 21.90 2 83.0 0.3460 0.7087cond:Hem 437.50 218.80 2 40559.0 3.4600 3.14E–02*cond:prof:group 1401.60 700.80 2 40574.0 11.0850 1.54E–05***

Table 4. (Continued)




Table 6. P600 in D: ANOVA result on the final linear mixed effect models for the mean amplitude between 800 and 1300 ms in D and A.


Ant:cond 1105.0 1105.00 1 40721.0 11.0600 8.83E–04***prof 259.0 259.00 1 87.0 2.5920 0.1110279group 3036.3 1518.20 2 85.0 15.1960 2.26E–06***Hem 984.6 492.30 2 40708.0 4.9280 0.007247**cond:prof 0.2 0.20 1 40726.0 0.0020 9.67E–01cond:group 7381.5 3690.80 2 40728.0 36.9430 < 2.2e–16***cond:Hem 2171.2 1085.60 2 40708.0 10.8660 1.91E–05***prof:Hem 1328.7 664.40 2 40708.0 6.650 0.0012954**Cen:cond 948.7 948.74 1 29044.5 13.2450 0.0002738***prof 693.8 693.78 1 83.9 9.6860 0.002539**group 968.6 484.29 2 83.1 6.7610 0.0019029**Hem 895.5 447.74 2 29033.2 6.2510 1.93E–03**cond:prof 2598.6 2598.60 1 29052.4 36.2780 1.73E–09***cond:group 4215.1 2107.55 2 29045.4 29.4230 1.72E–13***prof:group 79.8 39.88 2 83.0 0.5570 5.75E–01cond:Hem 115.9 57.96 2 29033.2 0.8090 0.4452665prof:Hem 424.2 212.11 2 29033.2 2.9610 0.0517703.cond:prof:group 647.4 323.69 2 29050.6 4.5190 0.0109089*cond:prof:Hem 1311.5 655.77 2 29033.2 9.1550 0.000106***Pos:cond 2659.2 2659.20 1 40718.0 37.9380 7.37E–10***prof 919.6 919.60 1 83.0 13.1200 0.0005031***group 485.6 242.80 2 83.0 3.4640 3.59E–02*Hem 1276.8 638.40 2 40706.0 9.1080 1.11E–04***cond:prof 6300.3 6300.30 1 40728.0 89.8850 < 2.2e–16***cond:group 3556.9 1778.40 2 40719.0 25.3720 9.72E–12***prof:group 111.4 55.70 2 83.0 0.7950 4.55E–01cond:Hem 2108.7 1054.40 2 40706.0 15.0420 2.95E–07***cond:prof:group 2012.2 1006.10 2 40724.0 14.3540 5.87E–07***



prosody-syntax integration in a second language

Documents